Facebook uses AI to accelerate discovery of new combinations of existing drugs

[ad_1]

Researchers from Facebook AI, the American company’s artificial intelligence research division, and Helmholtz Zentrum München (HMGU), a German health research center, have developed a model capable of accelerating the discovery of combinations effective drugs.

Their work can be viewed on bioRxiv but have not yet been peer-reviewed by a scientific journal. The model, dubbed “Compositional Perturbation Autoencoder” (CPA), is available in open source.

Fight against antibiotic resistanceScientists have looked at what are called “drug cocktails”, that is to say, the fact of re-adapting or combining existing molecules to find new treatments against cancer in particular. This method has two advantages: it makes it possible to fight against antibiotic resistance, one of the major health challenges according to the World Health Organization, and to reduce side effects.

However, finding the right “mixtures” turns out to be an extremely tedious task because it requires finding the right molecules, the right dosages, adequate planning and predicting their effects on the targeted cells, in particular on gene expression. This is the reason why researchers at Facebook and HMGU used a machine learning system capable of processing huge amounts of data.

Self-supervised model learningThey used self-supervised or self-encoding model learning, which avoids relying on labeled data that has long been essential to train models. This methodology is used in many research studies carried out by Facebook, explained Antoine Bordes, Co-managing Director of Facebook AI Research (FAIR), at The Digital Factory. In this case, it makes it possible to predict the effects of unknown combinations from data on known mixtures of drugs.

How model learning works relies on isolating key characteristics in a cell, such as the effects of a drug, combination, dosage, schedule, gene deletion, or of a specific cell type. Then, the model recombines these characteristics to predict their effects on the genetic expressions of the cell.

Scientists compare their work to dressing and undressing a cell. The different characteristics of a cell would be clothing elements, such as a hat, gloves or even a cap. During the learning phase, the model identifies the “dressed” cell then removes the clothing elements to dress it again in order to learn more about it. During the Review phase, the model can thus predict the best outfits, which corresponds to the most effective drug combinations.

A step towards ultra-personalized medicineThis work can be of interest to the pharmaceutical industry in order to more quickly select a drug combination to examine in order to find new remedies. In the long term, they pave the way for increased personalization of treatments according to the cells of each individual.

[ad_2]